---
mdfy_bundle: 1
id: ycb00hEg
title: "AI Memory Stack"
url: https://mdfy.app/b/ycb00hEg
document_count: 3
updated: 2026-05-09T20:22:44.329Z
source: "mdfy.app"
---
# AI Memory Stack

> Notes and references about AI memory systems — Mem0, Letta, OpenAI Memory, LLM Wiki, embeddings, vector DBs. Background reading for the mdfy direction.

## 1. Letta vs Mem0

*Source: https://mdfy.app/ycd01N9A*

> Side-by-side comparison with Letta — what they share, where they diverge.

# Letta vs Mem0

| | Mem0 | Letta |
|---|---|---|
| Storage | Vector + graph | Hierarchical (core / archival) |
| Authorship | Auto-extract | Auto-extract + manual edit |
| Hosted | Yes | Yes + self-host |
| Format | JSON memories | Block-based |
| OSS | Yes | Yes (formerly MemGPT) |

## Where they differ
Letta inherits MemGPT's "context windows are an OS resource" framing —
core memory (always loaded) vs archival (paged in). Mem0 is flatter:
all memories are vectors, top-K retrieved per turn.

## Where they're similar
Both treat memory as something the system *extracts from* the user's
behavior. Neither makes memory addressable as a URL the user can paste
into a different AI.


## 2. OpenAI Memory: Behind the Walled Garden

*Source: https://mdfy.app/ycd02zOQ*

> Why ChatGPT Memory is closed-by-design and what that costs users.

# OpenAI Memory: Behind the Walled Garden

ChatGPT's "Memory" feature is closed. Memories live in OpenAI's database,
shaped by their extraction model, only retrievable inside ChatGPT.

## What that means in practice
- Switch to Claude → memories don't follow
- Want to inspect what's stored → opaque
- Want to deploy "what ChatGPT knows about my project" to Cursor → can't

## The lock-in
Memory is the new vendor lock-in. Whoever owns your memory layer owns
the multi-AI workflow. mdfy bets on the opposite: memory you author,
URLs you paste anywhere.


## 3. mdfy Bundle Spec v1.0 (Draft)

*Source: https://mdfy.app/ycd04NxU*

> Our own spec — the open standard piece.

# mdfy Bundle Spec v1.0 (Draft)

A Bundle is a markdown document that wraps other markdown documents
as deployable AI context.

## Frontmatter (YAML)
```
---
mdfy_bundle: 1
id: <bundleId>
title: "..."
url: https://mdfy.app/b/<id>
document_count: N
updated: <ISO>
source: "mdfy.app"
---
```

## Body structure
- H1 = bundle title
- Optional blockquote = description
- H2 sections per included document, in sort order
- Each section: title, "Source: <url>", optional annotation, body

## Hub subtype
type: hub means body contains an INDEX (TOC + recent + categories)
not full doc bodies. Hub URLs reference doc URLs; AI follows links.

## Goal
Open standard. Any tool can produce or consume Bundles. mdfy.app is
the reference implementation.